AddSystemOneDecider

Answers every train's decisions through a typed decision model that speaks the System One request format (Jev, or another server that accepts it), by registering a SystemOneDecider as the IDecider, or, with a name, as a keyed SystemOneDecider only. For Nimble, use AddNimbleDecider, which fills in its model name and limits. See Typed decision models.

Signature

public static TBuilder AddSystemOneDecider<TBuilder>(
    this TBuilder configurationBuilder,
    Action<SystemOneOptions> configure
)
    where TBuilder : TraxEffectBuilder
 
public static TBuilder AddSystemOneDecider<TBuilder>(
    this TBuilder configurationBuilder,
    string name,
    Action<SystemOneOptions> configure
)
    where TBuilder : TraxEffectBuilder

SystemOneOptions

Returns

TBuilder, the same builder type that was passed in.

Example

services.AddTrax(trax => trax
    .AddEffects(effects => effects
        .UsePostgres(connectionString)
        .AddDecisionRecording()
        .AddSystemOneDecider(o =>
        {
            o.Endpoint = new Uri("https://api.typesafe.ai/v1/systemone");
            o.Model = "jev-1.13.0";
            o.ApiKey = configuration["Jev:ApiKey"];
        })
    )
);

Failures

SystemOneDecider implements IVetsQuestions, so what the model refuses whatever the state stops the host at startup instead of failing every run: more questions in one Decide than MaxQuestions, a question with blank instructions, fewer than two or more than MaxOptions options or levels, an option named twice, a kind of question the format cannot ask, and a state type JSON writes as a bare number, true or false. Each question is sent under its question key, the type's name without its namespace or the Key set on [Asks].

A choice or score answer without a confidence, which the format allows, takes the probability of the chosen option or the nearest level instead; one with neither is left out. A score whose probabilities are not keyed by every level from 0 is left out rather than shifted onto the wrong levels.

A retry waits for what Retry-After asks, plus a little jitter so a fleet of callers does not return at once, or else the doubling RetryDelay, jittered and capped at MaxRetryDelay. Cancelling the run cancels the request and is not retried. A failure's message gives the status, a few words on what it means, and the provider's request id from the x-typesafe-request-id response header when there is one; the request id is what the provider asks for when a failure is reported. The response body is never read into the message: it can echo the request, which carries the train's state, and the message is stored as the run's failure reason.

Every answer carries the model the response names. That is the name the request asked for, echoed back, not a version the server confirms, so pin the version where the model is deployed.

Remarks

• The options are checked when this is called, so unusable settings stop the host from starting.
• The unnamed form registers the decider as SystemOneDecider and as IDecider, and may be used once across this method and AddNimbleDecider; a second unnamed registration fails at startup instead of replacing the first.
• The named form registers a keyed SystemOneDecider under name and nothing else, so several models can sit side by side. Resolve each with GetRequiredKeyedService<SystemOneDecider>(name) and compose them, for example into a CascadingDecider registered as the IDecider; see Putting the model in front of a larger one. DecidedBy<TDecider>() resolves by type, so it cannot name a keyed decider.
• The decider is built on first use and owned by the container, which disposes its HTTP client when the host stops. A decision in flight when the decider is disposed keeps the outcome of its request.
• A SystemOneDecider built with its own HTTP client does not follow redirects. One built with an HttpClient you pass in uses that client, and a 3xx it hands back still fails the decision.

Package

dotnet add package Trax.Effect.Decisions.SystemOne